[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [dm-devel] SCSI Hardware Handler and slow failover with large number of LUNS

Chandra Seetharaman wrote:
Thanks for the response Mike.

On Mon, 2009-04-06 at 10:43 -0500, Mike Christie wrote:
Chandra Seetharaman wrote:
Hello All,

During testing with the latest SCSI DH Handler on a rdac storage, Babu
found that the failover time with 100+ luns takes about 15 minutes,
which is not good.

We found that the problem is due to the fact that we serialize activate
in dm on the work queue.

I thought we talked about this during the review?

Yes, we did and the results were compared to the virgin code (w.r.t rdac
handler) and the results were good (also I used only 49 luns) :

We can solve the problem in rdac handler in 2 ways
 1. batch up the activates (mode_selects) and send few of them.
 2. Do mode selects in async mode.
I think most of the ugliness in the original async mode was due to trying to use the REQ_BLOCK* path. With the scsi_dh_activate path, it should now be easier because in the send path we do not have to worry about queue locks being held and context.

little confused... we still are using REQ_TYPE_BLOCK_PC

But we only have one level of requests. I am talking about when we tried to send a request with REQ_BLOCK_LINUX_BLOCK to the module to tell it to send another request/s with REQ_TYPE_BLOCK_PC. Now we just have the callout and then like you said we can fire REQ_TYPE_BLOCK_PC reuqests from there.

I think when I wrote easier above, I meant to write a cleaner implementation.

I think we could just use blk_execute_rq_nowait to send the IO. Then we would have a workqueue/thread per something (maybe per dh module I thought), that would be queued/notified when the IO completed. The thread could then process the IO and handle the next stage if needed.

Why use the thread you might wonder? I think it fixes another issue with the original async mode, and makes it easier if the scsi_dh module has

can you elaborate the issue ?

I think people did not like the complexity of trying to send IO with soft irq context with spin locks held, then also having the extra REQ_BLOCK_LINUX_BLOCK layering.

to send more IO. When using the thread it would not have to worry about the queue_lock being held in the IO completion path and does not have to worry about being run from more restrictive contexts.

You think queue_lock contention is an issue ?

I agree with the restrictive context issue though.

So, your suggestion is to move everything to async ?

Do mean vs #1 or would you want to seperate and send some stuff async and synchronously?

Just wondering if anybody had seen the same problem in other storages
(EMC, HP and Alua).
They should all have the same problem.

Please share your experiences, so we can come up with a solution that
works for all hardware handlers.



dm-devel mailing list
dm-devel redhat com

To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo vger kernel org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]